Integrating domain knowledge into transformer for short-term wind power forecasting

Wind energy is an environmentally friendly source of energy and serves as an efficient supplement to conventional energy resources. Accurate wind power forecasting is crucial for effective decision-making in the daily operation of wind power plants. However, due to the heavy dependence on weather conditions, the variability and uncertainty associated with weather pose significant challenges to wind power forecasting. In this study, we propose a domain-knowledge integrated Transformer (DKFormer) model for short-term wind power forecasting. The proposed model integrates domain knowledge of wind power generation through three portable modules that play essential roles in data pre-processing, model training, and forecasting stages respectively. Additionally, by constructing boundary constraints that simultaneously utilize the data of both measured wind power and numerical weather prediction (NWP), the DKFormer model further reduces errors in multi-step wind power forecasting and improves overall forecast performance, particularly when input wind speed data exhibits dramatic variations. Furthermore, transfer learning techniques are employed to enhance the forecast capability of the DKFormer model using limited training data. Real-life datasets are used to evaluate the performance of the proposed DKFormer, demonstrating its superiority over conventional statistical models and DL models in short-term wind forecasting. Specifically, in day-ahead wind power forecasting experiments, our proposed DKFormer model achieves a 22.0% reduction in mean absolute error (MAE) while also exhibiting improved forecast stability compared to the conventional Transformer model.